fix(synthesis): keep the service worker alive during the reduce

[bryanmatthewsimonson]

Problem

Immediately after #196 (raising the reduce output cap to 32768), corpus synthesis started failing with "Synthesis failed: no response".

Root cause is a service-worker lifetime interaction the previous fix exposed:

• The reduce is the one long single fetch in the corpus flow (~350–455s for a large breadth brief; the 32768 cap makes it generate longer).
• The map phase survives because its many short xray:llm:corpus-map messages naturally reset the MV3 idle timer.
• After the last map lands, the reduce runs "cold" — no messages for minutes — so Chrome tears the background SW down mid-fetch. chrome.runtime.sendMessage's callback then fires with undefined (lastError = "message port closed"), which synthesis-block.js flattened to the opaque "no response."

Fix (mirrors the reader's proven pattern)

The reader already solves "long single LLM call in the SW" with startSwKeepalive() (src/reader/index.js ~2599, used by runQuickAudit). The corpus synthesis had no keepalive. This adds it:

• startSwKeepalive() in synthesis-block.js — a 20s setInterval ping to the zero-cost xray:llm:corpus-config handler, spanning the whole run (declared before the try, stop() in the finally so it's cleaned up on every exit path). Each onMessage delivery resets the idle timer, keeping the SW alive through the cold reduce.
• sendMessage() now reads chrome.runtime.lastError and resolves {ok:false, error, swLost:true} instead of a bare undefined — a lost SW surfaces a real reason plus an actionable "try again — the cached extracts make the retry cheap" hint, instead of "no response."
• CORPUS_REDUCE_TIMEOUT_MS 420000 → 480000 — with the SW now kept alive, the AbortController is the sole limiter, so give the largest breadth briefs headroom.

Verification

• 5-dimension adversarial review (keepalive lifecycle · sendMessage blast radius · MV3 semantics · render-guard interaction · timeout/abort), each finding independently verified → 0 confirmed, 0 uncertain (1 finding refuted as pre-existing/strictly-improved).
• npm run build clean · npm test 1976 pass / 0 fail · npm run lint 0 errors.

🤖 Generated with Claude Code